최근 포토로그


Caltech Pedestrian Dataset Computer Vision

헉 5기가가 넘는다...  그렇지만 둏은 자료군....



scrapped from : http://www.vision.caltech.edu/Image_Datasets/CaltechPedestrians/



Computational Vision at Caltech     |     MIS at TU Darmstadt

Caltech Pedestrian Dataset

peds01peds02peds03peds04

Description

The Caltech Pedestrian Dataset consists of approximately 10 hours of640x480 30Hz video taken from a vehicle driving through regular trafficin an urban environment. About 250,000 frames (in 137 approximatelyminute long segments) with a total of 350,000 bounding boxes and 2300unique pedestrians were annotated. The annotation includes temporalcorrespondence between bounding boxes and detailed occlusion labels.More information can be found in our CVPR09 paper.

Download

  • The training data for the Caltech Pedestrian Dataset is available here.There are six training sets (~1GB each), each consisting of 6-13 minutelong seq files, see the paper for more details. Detection results forall evaluated algorithms are also provided.
  • We are not releasing the testing data, please see "submitting results" below for information on how to include your trained pedestrian detector in the evaluation.
  • All videos are encoded using the seqfile format. An seq file is a series of concatenated image frames witha fixed size header, Matlab routines for reading/writing/manipulatingseq files can be found in Piotr's Matlab Toolbox (version 2.30 or later).
  • Associated Matlab code is available here.The annotations use a custom "video bounding box" (vbb) file format.The code also contains utilities to view seq files with the annotationsoverlayed, evaluation routines used to generate all the ROC plots inthe paper, and also the vbb labeling tool used to create the dataset (aslightly outdated video tutorial of the labeler is also available).
  • To allow for the exact reproduction of the INRIA ROC plots for full images, for convenience we are also posting the INRIA pedestrian full images/annotations in seq/vbb format as well as detection results for all evaluated algorithms.

Benchmark Results

Algorithm details and references can be found here. Note: some of the results below vary slightly from those in the CVPR09 paper due to simplified handling of ignore regions.
  1. Caltech Pedestrian Testing Dataset: All results inour CVPR09 paper were reported on this data (the data is not availablefor download, see submitting results for details). Results on 50-pixelor taller, unoccluded or partially occluded pedestrians are shown here, a more detailed breakdown of performance, as in the paper, can be found here.
  2. Caltech Pedestrian Training Dataset: Results on thetraining data (which is available for download). These results areprovided so researchers can compare their method without submitting aclassifier for full evaluation. Results on 50-pixel or taller,unoccluded or partially occluded pedestrians are shown here, a more detailed breakdown of performance can be found here.
  3. Caltech Pedestrian Japan Dataset: Similar to theCaltech Pedestrian Dataset (both in magnitude and annotation), exceptvideo was collected in Japan. We cannot release this data, however, wewill benchmark results to give a secondary evaluation of variousdetectors. Results on 50-pixel or taller, unoccluded or partiallyoccluded pedestrians are shown here, a more detailed breakdown of performance can be found here.
  4. INRIA Pedestrian Test Dataset: Results on the INRIA pedestrian full image data, obtained using the 288 positive test images (details given here). The ROC on the full image results is available here.
Last updated May 31, 2009.

Submitting Results

We are not releasing the test data at this time. Instead we askauthors to submit final, trained classifiers which we shall proceed toevaluate. We understand that this means additional effort for everyoneinvolved; however, our aim is to help prevent overfitting and to extendthe dataset's lifespan. Furthermore, it ensures that all algorithms areevaluated in precisely the same manner. Input/Output format: we areflexible since whatever IO format used, we simply write a wrapperfunction that allows for running the algorithm in a distributed manner.The only requirement is that the algorithm take in an image and returna bounding box and a score for each detection (as a matrix, text file,etc). The algorithm should perform multi-scale detection, detectingpedestrians at least 100 pixels tall (the returned detected boundingboxes can have additional padding) and performing any necessarynon-maximal suppression (nms). If need be nms and fast resampling code can be found in Piotr's Matlab Toolbox.Linux 32 or 64 bit binaries or Matlab code are ideal, Windows 32 bitbinaries are acceptable if need be. Finally, the algorithm should takea total of at most about 1 minute per 640x480 image (on a reasonablesingle core machine), with faster times being highly preferred, andmust be able to handle images as large as 1280x960. For algorithms thatutilize motion information, the input to the algorithm can be a pair ortriplet of images. For more sophisticated methods that require use ofthe entire video, we ask researchers to write routines that directlyutilize the seq files as input (using the provided seq support code).Please contact us if the above IO format is too restrictive for yourneeds.

Related Datasets

Below we list other pedestrian datasets,roughly in order of relevance and similarity to the Caltech Pedestriandataset. A more detailed comparison of the datasets (except the firsttwo) can be found in the paper.
  • Daimler:Also captured in an urban setting, update of the older DaimlerChryslerdataset. Contains tracking information and a large number of labeledbounding boxes.
  • NICTA:A large scale urban dataset collected in multiple cities/countries. Nomotion/tracking information, but signficant number of uniquepedestrians.
  • ETH: Urban dataset captured from a stereo rig mounted on a stroller.
  • INRIA: Currently one of the most popular static pedestrian detection datasets.
  • PASCAL: Static object dataset with diverse object views and poses.
  • USC: A number of fairly small pedestrian datasets taken largely from surveillance video.
  • CVC: A fairly small scale urban pedestrian dataset.
  • MIT: One of the first pedestrian datasets, fairly small and relatively well solved at this point.


덧글

댓글 입력 영역